Evolution and Rawlsian social choice in matching

GAICIL Insights

Can a market be fair without anyone designing it to be fair?¹ Evolution and Rawlsian social choice in matching, by Ennio Bilancini, Leonardo Boncinelli and Jonathan Newton, finds a surprising yes, under a specific behavioral condition. The paper studies the marriage problem under dynamic rematching and asks which matchings emerge in the long run as players rematch over time.

The answer carries an ethical signature. If players with higher payoffs are less likely to experiment with non-best response behavior, then the long-run matchings belong to the set of Rawlsian stable matchings, those maximizing the payoff of the worst-off player. Alternative behavioral rules fail to select Rawlsian stable matchings in some environments. The authors describe the result as an evolutionary axiomatization of Rawlsian stable matchings: in marriage problems under perturbed pairwise best response dynamics, Rawlsian stable matchings emerge as long-run social norms if and only if players follow condition dependent behavioral rules.

The intuition is lovely. When the well-off experiment less and the badly-off experiment more, rematching keeps searching for improvements for the worst-off until no further gains exist. Concern for the least advantaged emerges from individual behavior, with no planner in sight. Restlessness at the bottom and contentment at the top produce justice as a by-product.

Now the AI angle. Artificial intelligence now runs many matching markets: workers and jobs, users and services, students and programs. The behavioral rules embedded in these algorithms, including how they experiment and respond to outcomes, influence which matchings emerge. The paper suggests that rules conditioning experimentation on current payoffs can steer outcomes toward matchings that favor the worst-off. Fairness becomes a design parameter.

That yields a concrete design principle for AI-driven matching platforms. Platforms routinely experiment, testing alternative matches to learn preferences. Concentrating experimentation on participants with poor current matches, while leaving well-matched participants stable, could improve outcomes for those who need it most. The theory connects this design choice to long-run distributional outcomes. A platform's exploration budget, spent at the bottom, is a redistribution policy that costs nothing in efficiency terms.

The authors list future directions: extending the result to the roommate problem, one-sided matching or the assignment problem with transferable utility; behavioral axiomatizations of other social choice functions; and, following an approach pioneered by Sandholm for population games, designing institutional features of the marriage market to facilitate desired norms, including marriage costs, divorce costs and restrictions on admissible partnerships due to geography and social status.

That last direction maps directly onto AI-era policy. Switching costs, the geographic scope of markets and eligibility restrictions shape outcomes. AI platforms can cut switching costs and widen geographic scope, potentially changing the long-run norms that emerge. The framework offers a way to analyze whether such changes move outcomes toward or away from Rawlsian matchings, a question regulators of labor and education platforms should be asking.

The paper also reflects a broader market-design theme: decentralized processes can deliver socially desirable outcomes when behavioral rules have the right properties. Central assignment of matches becomes unnecessary when rematching dynamics, guided by suitable rules, converge to fair and stable outcomes. That supports confidence in decentralized markets enhanced by well-designed algorithms, provided designers pay attention to who experiments and when.

Taken together, fairness can emerge from the bottom up. When behavior conditions experimentation on outcomes, long-run matchings protect the worst-off. As AI reshapes matching markets across the economy, designers have the chance to embed rules that produce such outcomes. Theory of this kind shows how individual behavior and social justice can align, and how innovation in market design can make matching work better for everyone.

Something for platform engineers to sit with: the fairest algorithm may be the one that keeps searching hardest on behalf of those it has served worst so far.